AI for lawyers: 5 everyday tasks to delegate, prompts included

Summarising a forty-page judgment, rebuilding a timeline from a binder of exhibits, re-reading a contract for the third time: part of every lawyer's day goes into tasks that require neither your training nor your judgment. AI for lawyers has moved from gadget to reflex: 74% of lawyers and in-house counsel say they used it regularly in 2025, up from 53% a year earlier, according to the annual study by legal publisher Lamy Liaisons.

The question that matters is practical: what exactly should you delegate, and how do you phrase the request to get a usable result? We selected five everyday tasks that lend themselves well to a first pass by artificial intelligence. For each one: a prompt ready to copy, and the points to check before you take over.

A word of honesty first. We will not promise you "30 minutes saved per judgment": nobody has measured that seriously. The only published figure is a global one: 62% of legal professionals report recovering between 6 and 20% of their working week thanks to these tools, according to the Wolters Kluwer "Future Ready Lawyer" 2026 survey.

AI for lawyers does the first pass, you sign

The division of labour fits in one sentence: the tool produces a first version of documentary work, the lawyer keeps the review, the legal qualification and the signature. The guidelines published by AVOCATS.BE (the association of Belgium's French- and German-speaking bars) on 31 January 2025 set that frame without banning much: they classify the tools, set security conditions and recall that professional liability cannot be delegated.

A task is fit for delegation when three conditions are met: the raw material sits in your documents, the format of the deliverable can be described in a few lines, and the result can be checked faster than it can be produced. The five tasks in this article tick all three boxes.

Prerequisites before you start

• Digitised, readable documents: a crooked scan produces a crooked timeline.
• A tool whose data flows you actually know: hosting location, sub-processors, data processing agreement (DPA), GDPR compliance. For documents covered by professional secrecy, a free consumer chatbot does not answer those questions.
• Your firm's sign-off: the French-speaking Bar of Brussels offers a model charter for AI use in law firms, useful even for a solo practice.
• The verification reflex: no reference leaves the office without being checked against its source.
• A quiet hour to test the prompts below on a closed file, not on tomorrow's emergency.

The delegation circuit in one diagram

flowchart TD
    A["Repetitive task identified"] --> B{"Is the raw material<br/>in your documents?"}
    B -- "No" --> C["Do not delegate yet:<br/>gather the exhibits first"]
    B -- "Yes" --> D["Structured prompt:<br/>context, format, guardrails"]
    D --> E["First pass by the AI"]
    E --> F{"Lawyer's check:<br/>sources cited and verified?"}
    F -- "No" --> G["Prompt corrected<br/>or manual takeover"]
    G --> D
    F -- "Yes" --> H["Deliverable reviewed and<br/>signed by the lawyer"]

The checkpoint is not decorative: it is what turns plausible text into a lawyer's work product.

1. Summarising a long judgment

A thirty-page judgment can be summarised in five imposed points: court and references, essential facts, legal question, the court's position, practical takeaway. Summarising legal documents is the most widespread use among legal professionals: 74% cite it in the Lamy Liaisons study.

Summarise the judgment provided in 300 words maximum,
in five points:
1. Court, date, docket number or ECLI reference
2. Essential facts (three sentences maximum)
3. Legal question(s) raised
4. The court's position and decisive grounds
5. Practical takeaway for a similar case

For each point, quote the exact passage of the judgment
supporting it. If an element does not appear in the document,
write "not specified" instead of inferring it.

That last instruction separates a serious tool from a plausibility generator: without it, the AI fills the gaps. Two things remain for you to check: that the quoted grounds are indeed the ones carrying the decision (models sometimes overweight an obiter dictum), and that point 5 matches your own reading. The practical lesson of a decision remains a lawyer's act.

2. Rebuilding the case timeline

A reliable timeline is a three-column table: the date, the fact, the exhibit that establishes it. It is the most thankless task in a bulky file, and one of the most mechanical: exactly the profile of a good delegation.

From the exhibits provided, build the timeline of the case.
Format: three-column table (date, fact, exhibit and page).
Rules:
- Strictly chronological order, one row per dated event
- For each row, give the precise reference of the exhibit
- If two exhibits contradict each other on a date, flag the
  conflict instead of settling it
- List separately the events mentioned but not dated

The two final guardrails make the value of the output. Flagged date conflicts show you where to dig; the list of undated events keeps you from believing the timeline is complete when it is not. And the "exhibit" column turns the table into a hearing tool: every row points back to its source.

3. Hunting down inconsistencies in a contract

A consistency review crosses every clause with all the others: cross-references, definitions, amounts, deadlines, obligations. Pure attention work, with no legal subtlety, and legal professionals already delegate it widely: 69% cite contract review and adjustment among their AI uses (Lamy Liaisons study).

Analyse this draft contract and flag only internal
inconsistencies:
1. Clauses that contradict each other
2. Cross-references to non-existent or misnumbered articles
3. Terms used but never defined
4. Obligations whose debtor is not clearly designated
5. Amounts, dates or deadlines that differ from one clause
   to another

For each inconsistency, quote the clauses concerned word
for word. Do not propose any rewording at this stage.

The ban on rewording is deliberate: diagnosis first, pen second. The AI flags formal contradictions; the balance of a clause, its advisability and its validity remain your analysis.

4. Preparing case law research

AI-assisted case law research only has value if every decision cited exists and opens in one click. Identifying case law ranks third among uses (55% in the Lamy Liaisons study), and the Belgian volume explains why: JUPORTAL publishes around 145,000 decisions going back to 1990, the Council of State counts some 220,000 since 1994, and the Council for Alien Law Litigation has issued around 268,000 rulings since 2007. Nobody sweeps through that manually.

Search the sources for relevant case law on
[precise legal question, with the field and the period].
For each decision retained:
- Court, date, docket number or ECLI reference
- Principle established (two sentences maximum)
- Relevant passage quoted verbatim

Only retain decisions whose text you can quote.
If you find nothing relevant enough, say so.

This prompt assumes a tool connected to real databases. This is where an assistant built on official sources parts ways with a generic chatbot: Jef, for instance, continuously indexes JUPORTAL, the Council of State, the Council for Alien Law Litigation and EUR-Lex, and every answer cites the passages used (the full catalogue is at jef.chat/en/sources). On a generic model with no documentary grounding, the same prompt produces the illusion of research. We come back to it below: the bill can be steep.

5. Drafting the first version of a formal notice

A formal notice (mise en demeure) follows a stable structure: reminder of the facts, contractual or legal basis, quantified demand, deadline, reservation of rights. That repetitive character makes it the ideal candidate for an automatic first draft, like the firm's other recurring letters (acknowledgements of receipt, requests for exhibits, deadline reminders).

Draft a first version of a formal notice (mise en demeure)
under Belgian law.
Context: [debt or obligation, amount, due dates, reminders
already sent, available exhibits]
Structure:
- Reminder of the facts and the contractual or legal basis
- Precise demand: principal amount, interest, deadline of
  [8 days or other]
- Reservation of all rights and remedies
Tone: firm and sober, without rhetoric.
Place in square brackets every element to verify or complete
before sending.

Re-read three things before signing: the proposed legal bases (a statute cited from memory by a generic model is a hypothesis, not a basis), the amounts and dates, and the tone. The brackets requested in the prompt materialise what awaits your validation: an honest draft shows its gaps.

The five tasks in one table

The trap: case law that does not exist

Symptom. Submissions that look impeccable, supported by decisions with plausible references... that nobody can find. Belgian courts and tribunals have recorded at least four abusive uses of AI by lawyers, according to AVOCATS.BE. The Antwerp Court of Appeal settled the matter in 2025: a €2,500 fine for manifest abuse of process and €7,500 in damages per opposing party (there were three), €25,000 in total, for submissions resting on non-existent case law and legal sources invented out of thin air. On 15 December, the Ghent Enterprise Court in turn received submissions citing rulings of the Court of Cassation and the Constitutional Court that never existed.

Diagnosis. A generalist generative model consults no database: it predicts the most plausible text. A reference produced without documentary grounding is a statistical prediction dressed up as a source. The problem is not a "bug" waiting for a patch, it is the nature of the tool: asking a database-less chatbot for case law is like asking for an address from someone who invents streets in a valid format.

Fix. Two moves, in this order. Restrict research to tools that quote the exact passage and give access to the source; then verify every reference in the original database (JUPORTAL for Belgian case law) before it enters a procedural document. Count one to two minutes per reference. The hearing on abuse of process lasts considerably longer.

What AI for lawyers will not replace

Three things are never delegated, and we stand by the position: the legal qualification of a case, the strategic call (settle, plead, play for time), and the review that precedes the signature. An AI compares, sorts, summarises, drafts. It does not appear in court and does not wear the robe: professional liability remains yours.

The gain therefore does not lie in legal work disappearing. It lies in time shifting: fewer hours searching, compiling and formatting, more hours analysing and deciding. The Wolters Kluwer survey calls it an "80/20 reversal": collection gives way to analysis.

Start small. Pick one of the five tasks, test the prompt on a closed file, measure your own verification time, adjust. AI for lawyers should be judged on evidence, like everything else. Jef (https://www.jef.chat/en) applies that logic: the assistant turns your firm's document base into a conversational agent with sourced answers, built on official Belgian and European sources, in a partitioned, encrypted workspace. Registration is free, with no commitment and 50 messages included: enough to test the prompts in this article on your own documents.